The Unfireable Safety Kernel: Execution-Time AI Alignment for AI Agents and Other Escapable AI Systems — ThinkLLM